Meta stumbles, Anthropic expands, OpenAI plays politics

Zuckerberg admits AI agents are behind schedule, Anthropic launches drug discovery and a Samsung chip deal, and OpenAI floats a 5% government stake.

The big picture

The AI industry’s internal cracks are showing: a CEO admitting his agents are behind, a lab pivoting into drug development, and the dominant commercial player trying to buy political goodwill with equity. This batch is less about new capabilities and more about where the money and power actually flow.

Meta’s big year is running a little late

At an internal town hall, Zuckerberg acknowledged that Meta’s AI agent push is behind his own timeline, while his AI chief offered a sunnier read of the same facts. That gap between CEO honesty and executive spin is worth watching. Meta reorganized significant chunks of the company around agents — so “slower than planned” has real organizational consequences, not just PR ones. (TechCrunch, The Decoder)

Separately, Meta quietly launched Pocket, an app for generating playable mini-games from text prompts. (TechCrunch) Vibe-coded games are a fun demo format, but launching it quietly while admitting your agent strategy is struggling feels like a shiny object to distract from the headline.

Meta is also working on a cloud infrastructure business to sell excess compute — going head-to-head with AWS, Google Cloud, and Azure. (TechCrunch) The SpaceX comparison (turning internal capability into an external revenue line) is apt, but Meta is entering a market where the incumbents have a decade of enterprise relationships.

Anthropic is doing a lot at once

Anthropoc is reportedly in talks with Samsung to develop a custom AI chip, following OpenAI’s Broadcom partnership by about a week. (TechCrunch) Every top-tier lab is now pursuing silicon independence — the days of simply buying Nvidia capacity and calling it a strategy are over.

More surprising: Anthropic announced it will develop its own drugs for neglected diseases that pharma considers unprofitable, alongside Claude Science, a research workbench that unifies tools, datasets, and figure generation for scientists. (The Decoder, The Verge) Novartis’s CEO cited AI potentially cutting drug development from 12 years to 7-8 and doubling success rates from 8% to 16%. Those numbers are aspirational, but Anthropic going from AI safety research to owning drug programs is a genuine strategic leap — and one that will invite significant regulatory scrutiny.

Also: Claude Fable 5 is back. After weeks of negotiating with the Trump administration over export controls, the Department of Commerce lifted restrictions and Anthropic is restoring global access. (The Verge) The fact that a model required weeks of government negotiation to re-enable is a preview of how tangled AI geopolitics has become.

OpenAI’s political and philosophical bets

OpenAI floated giving the U.S. government a 5% equity stake, worth roughly $42 billion at the company’s current $852 billion valuation, reportedly pitched to Trump last year and now surfacing publicly. (TechCrunch, The Verge) The framing as “sharing the upside of AI” is generous; the subtext is regulatory protection. Whether a sovereign wealth fund stake changes anything about how OpenAI is regulated is the real question.

Cofounder Greg Brockman is pitching a future where software has almost no interface and people stop learning apps entirely — an always-on, context-aware agent handles everything. (The Decoder) He even admitted ChatGPT plugins failed because the models weren’t ready. Honest in hindsight, but that vision remains very far from what Codex and Operator can actually do today.

Quick hits

Microsoft launched an AI deployment subsidiary backed by a $2.5 billion commitment, joining Amazon, OpenAI, and Anthropic in standing up dedicated deployment units. (TechCrunch)
Jersey Mike’s IPO filing apparently mentions AI — a useful canary for how saturated the term has become in corporate filings. (TechCrunch)